Abstract
In industrial settings, safety-critical systems generate non-uniform, non-stationary data streams that require timely and accurate monitoring to prevent catastrophic failure. This paper proposes a novel change-point detection (CPD) method based on Online Dynamic Mode Decomposition with control (ODMDwC). By tracking linear approximations of non-linear systems while incorporating control effects, the method dynamically adapts to system aging and seasonality. Our approach enables the detection of shifts in spatial, temporal, and spectral patterns, preserving the correspondence between the anomaly score and the physical extent of change. Validated on synthetic and real-world benchmarks, including the Skoltech Anomaly Benchmark (SKAB) and industrial battery storage systems, our method demonstrates superior accuracy and computational efficiency compared to traditional SVD-based and deep learning approaches.
Introduction
In high-stakes industrial environments—such as power plants, chemical reactors, or massive battery arrays—detecting a change in system behavior is a safety imperative. When a cooling fan fails or a valve begins to stick, operators require real-time alerts. Traditional monitoring often relies on offline models that cannot adapt to the "new normal" of an aging system, or simple threshold-based SCADA systems that are prone to false alarms.
Industrial data streams present unique challenges: they are non-stationary, highly correlated in time (non-i.i.d.), and influenced by deliberate control commands. A shift in data might indicate a fault, or it might simply be the system responding correctly to a change in input. To bridge this gap, we introduce ODMD-CPD, a method that learns the underlying "rhythm" of a system on the fly and distinguishes between healthy responses and genuine dynamic shifts.
Methodology
From SVD to DMD
The core of our approach is Dynamic Mode Decomposition (DMD). While traditional methods like Singular Value Decomposition (SVD) focus on spatial correlations—how sensors move together—DMD captures temporal dynamics. It decomposes signals into modes characterized by specific frequencies and growth or decay rates. This allows the model to monitor the physical behavior of the system rather than just signal magnitude.
Online Updates and Control
To make DMD suitable for real-time industrial use, we implement several key innovations:
- Online Updates: Using a recursive least-squares formulation, the DMD matrix is updated as new data arrives. A sliding window allows the model to "forget" old data, adapting to gradual changes like equipment wear and tear.
- Incorporating Control (DMDc): The algorithm augments the state matrix with control inputs. This allows the model to "subtract" the expected effect of a control action (e.g., a pressure rise following a valve opening) from the observed data, preventing false positives triggered by routine operations.
- Truncated Online SVD: To handle noise and high-dimensional data efficiently, we utilize a modified Brand/Zhang algorithm to maintain a stable, low-rank computational environment.
- Hankelization: By using time-delay embeddings (stacking past observations), we capture non-linearities and broad-band features that a single data snapshot might miss.
Detection Logic
The algorithm maintains two windows: a base window (the reference state) and a test window (the most recent data). Both are projected onto the current DMD subspace. If the test window exhibits a significantly higher reconstruction error than the base window, a change-point is flagged, indicating that the system's underlying physics have shifted.
Experimental Results
Benchmarking Performance
We validated ODMD-CPD across diverse datasets, ranging from synthetic steps to complex industrial benchmarks:
- SKAB (Skoltech Anomaly Benchmark): On this water circulation system benchmark, ODMD-CPD achieved a Numenta Anomaly Benchmark (NAB) score of 34.29, outperforming the state-of-the-art deep learning model MSCRED (32.42) and significantly beating Vanilla LSTMs and Isolation Forests.
- CATS (Controlled Anomalies Time Series): In a test involving 5 million snapshots of a complex dynamical system, our method achieved results comparable to deep learning but was 24x faster, making it viable for true 1Hz real-time deployment.
- BESS (Battery Energy Storage System): In a real-world case study involving a faulty HVAC cooling fan, the method successfully detected the failure and provided "precursor" warnings before the fault became critical.
Sensitivity and Robustness
In synthetic tests, the DMD-based score proved much cleaner and less noisy than SVD-based references. Because DMD focuses on frequency and growth rates (the "energy" of the dynamics) rather than just signal energy, it can detect minor shifts that are invisible to traditional subspace methods.
Discussion and Conclusion
ODMD-CPD offers a powerful combination of physics-informed modeling and computational efficiency. By explicitly modeling control inputs, it avoids the common pitfall of false alarms during operational changes. Furthermore, because DMD modes have physical interpretations (frequency and damping), the method provides a level of interpretability often missing in "black box" deep learning models; operators can potentially see which part of the system dynamics changed.
While the method requires careful selection of hyperparameters—such as window sizes and the number of time-delays—it represents a significant step forward for industrial health monitoring. Future work will focus on universal normalization of the CPD statistic to simplify threshold setting across different types of industrial equipment. Ultimately, ODMD-CPD provides a robust, real-time solution for protecting both profit and life in safety-critical environments.
Summary updated